Déjà-vu, evolution and revolution: artificial intelligence in tax

In the tax departments of larger companies, at many tax consultants and at tax consulting firms, this year could be remembered as the starting point of a revolution. Hardly a month goes by without new developments in the field of automated consulting services being presented at conferences and in relevant professional journals. Nobody doubts that new technologies will change the work of tax departments and the professional profile of tax consultants. But are the new technologies really revolutionary?

Human intelligence is characterised by the fact that knowledge that has remained unused for a long time fades away. This can be an advantage. Not only because it corrects undesirable developments, but also because new developments are more attractive and motivating than old ones. Nevertheless, when it comes to artificial intelligence in the tax field, some people may experience déjà vu. Haven’t there been prototypes of artificial intelligence in the tax field for a long time?

A look at the old volumes of the magazine Datenverarbeitung – Steuern – Wirtschaft – Recht (DSWR), the organ of DATEV eG that was discontinued in 2006, provides certainty. In 1984, the question was already posed to the readership of tax consultants as to how much intelligence should be granted to computer systems. An article in the same magazine from 1989 impresses with an exuberant footnote apparatus – and gives the impression that Germany was ahead in research (“Aim of AI: Simulation of human problem solving”). The theory of the 1970s was followed by a “hype” of expert systems in the 1980s. Quite a few prototypes were developed and tested in practice. Systems such as BILEX, GUVEX, FINEX and INCOSS would today be marketed under the banner of data science. Experiments have also been carried out with systems for the automated generation of balance and audit reports. Computer-aided legal expert systems (S. Grundmann, DSWR 1987, 213) were also intensively discussed and widely researched at an equally early stage. This historically impressive discourse and practical testing came to an end in the early 1990s – at about the same time as the triumphant advance of Internet applications began. Conspiracy theorists believe that the professions have decreed a halt to development. It is more likely that priorities in both jurisprudence and computer science had shifted with the challenges of the Internet age.

The prototypes that have been presented recently in Berlin at an event on artificial intelligence in the tax field by the tax consulting firm WTS together with the German Research Center for Artificial Intelligence (DFKI) (see for example the report of the DStV and here in the blog) tie in with the expectations and goals of the 1980s. However, they rely on a different form of artificial intelligence than the expert systems of that time. This technical evolution could not only change business models, but also have repercussions on the development of law.

Artificial intelligence can be implemented for the representation of knowledge and the simulation of human problem solving in two different ways, rule-based or case-based. Case-oriented systems can carry out pattern matching or inductively develop decision trees by learning from cases. The early expert systems were predominantly rule-based. The necessary technology was available quasi with the discovery of computer technology. The process of knowledge implementation by programmed if-then-sentences was, however, hardly suitable for large and dynamic control systems. The rules must be translated into program code by human experts and the program code must be continuously adapted to changes in legal reality. This effort is enormous. For this reason, technology could only establish itself in niches and mass markets. The question-answer systems that can be found in commercially available tax return assistance systems basically work with this technology. Today they are additionally equipped with practical document recognition modules. However, much-discussed new providers fail in some cases because of the simple complexity of a joint investment. Only machine-readable (tax) law could allow a renaissance of rule-based expert systems.

The alternative model of case-based expert systems presupposes sufficiently large collections of cases containing facts and associated legal conclusions. From this, suitable methods can be used to either – inductively – derive own rule systems or to directly determine a similar case in the database for a new case in order to then transfer the conclusions linked in the comparative case to the new case. Although the necessary methods of machine learning

by Prof. Dr. Heribert M. Anzinger